CN102706881A - Cloth defect detecting method based on machine vision - Google Patents

Abstract

The invention belongs to the technical field of image processing and pattern recognition, and relates to a cloth defect detecting method based on machine vision. The cloth defect detecting method includes performing power spectral density analysis for an image of a normal cloth texture and acquiring central frequency F and an azimuthal angle theta of the texture; constructing an SXL adaptive Gabor filter bank; filtering the image of the normal cloth texture to obtain a feature image group, and computing the mean value and the variance of each image of the feature image group; acquiring an image of to-be-detected cloth; filtering the image of the to-be-detected cloth to obtain a feature image group; performing threshold post-processing for the feature image group of the image of the to-be-detected cloth to obtain an absolute feature image group; carrying out normalization processing; fusing images and performing binarization processing for the images to obtain a detected binary image; and removing noise interference to obtain a final detection result. The S represents the number of the selected central frequency, and the L represents the number of the selected azimuthal angle. The cloth defect detecting method has the advantages of high universality and efficiency.

Description

Cloth flaw detection method based on machine vision

Technical field

The invention belongs to the image processing and pattern recognition field, relate to a kind of cloth flaw detection method based on machine vision.

Background technology

In textile industry, cloth Defect Detection link is the key factor that influences product quality.In process of producing product, because the influence of factors such as machinery and equipment operating mistake or mechanical disorder contains various flaws usually in the cloth.At present, domestic most of weaving mills mainly rely on the artificial vision to detect the cloth flaw, have open defect: (1) labour intensity is big, and detection speed is slow, and detection efficiency is low; (2) testing staff need concentrate on work for a long time, and whole work is healthy unfavorable to the workman; (3) influence of testing staff's subjective factor is big, and false drop rate and loss are high, generally can only detect 40%～60% flaw.

Mainly contain the online perching of the Cyclops system of Belgian Barco company in the market, Fabriscan automatic Cloth Inspecting System and the online cloth checking system that German Opdix photoelectron technology company develops of I-TEX perching system and the Switzerland Uster company of little looking (EVS) company liked by Israel.Hong Kong University has developed the cloth Defect Detection CAVIS of system based on the Gabor bank of filters at home, has certain verification and measurement ratio.The detection system price of more than introducing is very expensive; Detection system all has been used for actual cloth production testing process; And their detection effect is general and have a very strong limitation; They all require cloth must be that single background or texture are isotropy, and for the flaw flaw of less energy, is difficult to correct detecting.

Summary of the invention

In order to overcome the deficiency of cloth flaw manual detection; Big, the slow shortcoming of detection speed of cloth Defect Detection algorithm computation amount to based on the Gabor bank of filters, the present invention provides a kind of highly versatile, cost the low cloth flaw detection method based on machine vision that also can improve detection efficiency and effect.Technical scheme of the present invention is following:

A kind of cloth flaw detection method based on machine vision comprises the following steps:

(1) the normal cloth texture image of industrial camera collection P (x, y);

(2) (x y) utilizes power spectral-density analysis, obtains the centre frequency F and the azimuth angle theta of cloth image texture to P;

(3) two groups of parameters of definite figure image texture: the centre frequency group (F, F/2); Position angle group (θ, θ+pi/2) is utilized these two groups of parameter settings and is formed the Adaptive Gabor bank of filters of S * L, wherein, and the centre frequency number that the S representative is chosen, the position angle number that the L representative is chosen;

(4) under off-line case, (x y) carries out the filtering of Adaptive Gabor bank of filters, obtains characteristic image group H to normal cloth texture image P _{M, n}(x y), calculates the average (Mea of its every width of cloth figure _{M, n}) and variance (Std _{M, n}), as the threshold value that detects visual flaw.

(5) gather cloth image I to be detected (x, y);

(6) (x y) carries out filtering, obtains characteristic image group H to the cloth image I with the Adaptive Gabor bank of filters that makes up _{M, n}(x, y);

(7) use the average (Mea that obtains from (4) _{M, n}) and variance (Std _{M, n}) to H _{M, n}(x y) carries out the threshold value aftertreatment and obtains the absolute feature image sets

Wherein, τ is adjustable parameter, span 2-3;

(8) to image sets F _{M, n}(x y) carries out the normalization processing and obtains relative characteristic image sets N _{M, n}(x, y);

(9) to image sets N _{M, n}(x, the y) fusion of travel direction angle and frequency, and carry out binary conversion treatment and obtain detecting bianry image I _Bm(x, y);

(10) utilize the area information of minimum flaw to remove bianry image I _Bm(x, the noise in y) obtain final detection result I _Out(x, y);

(11) judge testing result I _Out(whether x y) contains flaw.

As preferred implementation, described cloth flaw detection method based on machine vision, in the step (9), the fusion method of using the Bernoulli criterion is to image sets N _{M, n}(x, y) fusion of travel direction angle and frequency obtains image.

Cloth Defect Detection involved in the present invention technology is at first utilized the power spectrumanalysis of normal cloth texture image, obtains the characteristics such as centre frequency and deflection of cloth texture, utilizes these characteristic informations to set the parameter of adaptive Gabor bank of filters; Then, utilize the Adaptive Gabor bank of filters that normal cloth texture image is carried out filtering, to the threshold value of characteristics such as filtered its average of image extraction, variance as online cloth Defect Detection; Equally, utilize the Adaptive Gabor bank of filters that the cloth sample image is carried out filtering, and the threshold value of utilizing off-line to obtain is carried out the detection of cloth flaw; At last, the characteristic image after the threshold value is carried out processing backs such as normalization, image co-registration, binaryzation, denoising flaw is judged, output flaw information, the whole process of completion cloth Defect Detection.The cloth flaw detection method that the present invention proposes; This technology has been set the parameter of adaptive Gabor bank of filters with the power spectrumanalysis of cloth texture image; Effectively reduced the number of wave filter, and improved and detected effect, practice thrift human cost, improved the confidence level of detection efficiency and product and reduced loss and false drop rate significantly.Testing result as shown in Figure 5 shows that the Gabor bank of filters of designing can detect various types of flaws.The present invention can detect 80% cloth flaw, and verification and measurement ratio reaches 85%.

Description of drawings

Fig. 1: the FB(flow block) of cloth Defect Detection algorithm.

Fig. 2: Adaptive Gabor bank of filters parameter setting block diagram.

Fig. 3: several kinds of common cloth flaw images: (a) disconnected flaw image, (b) face grain flaw image, (c) the weft crackiness flaw image knitted; (d) greasy dirt flaw image (e) is slightly tied the flaw image, (f) colo(u)r streak flaw image; (g) hand pick flaw image, (h) broken hole flaw image, (i) skips flaw image; (j) ring flaw image, (k) double ends flaw image, (l) double weft flaw image.

Fig. 4: the disconnected image of knitting flaw image and the generation of whole testing process thereof.Image (A) is the disconnected flaw image of knitting; Image (A1) and (A2) be normalized horizontal properties result after the Gabor filtering; Image (A3) is that (A4) is normalized vertical characteristic result after the Gabor filtering; Image (A5) is centre frequency F ₁On fusion results; Image (A6) centre frequency F ₂On fusion results; Image (A7) is total fusion results; Image (A8) is the result after the binary conversion treatment.

Fig. 5: flaw image and testing result that some are common.In Fig. 5, from (5-1) to (5-11) is respectively face grain, weft crackiness, greasy dirt, thick knot, colo(u)r streak, hand pick, broken hole, skips, ring, double ends, double weft flaw and their testing results.In every pair of image, left side figure is the former figure of flaw image, and right figure is testing result figure.

Embodiment

The general flow chart of the cloth Defect Detection technical method based on machine vision of the present invention is as shown in Figure 1; At first from normal cloth texture image is carried out power spectrumanalysis; The textural characteristics information that utilization is obtained is set (can referring to Fig. 2) to the Adaptive Gabor bank of filters; Through the cloth sample image is carried out filtering, extract the judgement of characteristic and flaw information and carry out the detection of cloth flaw.

Referring to Fig. 3, common cloth flaw for disconnected knit, face grain, weft crackiness, greasy dirt, thick knot, colo(u)r streak, hand pick, broken hole, skips, ring, double ends, double weft flaw etc.Utilize below the disconnected whole testing process of knitting image of Fig. 4 is done further detailed description.

1. cloth texture image power spectrumanalysis

The image power spectrum is expressed as the distribution of picture power on frequency, and picture power is defined as:

P(u，v)＝|F(u，v)| ²＝R ²(u，v)+I ²(u，v)

Wherein, (u v) is the energy size of image to P, and (u v) is the real part behind the visual Fourier transform to R, and (u v) is the imaginary part behind the visual Fourier transform to I.(u v) is the image pixel point coordinate.

Phasing degree definition after the conversion:

Utilize the image power spectrum can obtain the centre frequency of image texture and the size of deflection.Utilize these two parameters can set the Adaptive Gabor wave filter that matees with the cloth texture.

The Gabor bank of filters is complete but has redundantly that the filtered that is to say the Gabor bank of filters is the reconstruct original signal fully, but is not quadrature between the wave filter.Set a highest frequency F _Max, the centre frequency of bank of filters is elected F as _Max, 2 ^-1F _Max, 2 ^-2F _Max... make that like this whole filter group can well cover the frequency domain scope of trying one's best big and make that the correlativity between the wave filter is very little.The less flaw of most sizes all appears at follows on the consistent or perpendicular direction of main grain direction; Therefore the position angle of wave filter is only selected get final product with the main grain direction one vertical both direction angle of making peace, and the position angle of increase wave filter means the increase calculated amount.

2. the Adaptive Gabor bank of filters is set

(1) two-dimensional Gabor filter

The Gabor conversion has been widely used in visual texture analysis, recognition of face, target detection and identification and other fields.In fact, the kernel function of Gabor conversion is a Gaussian function, because the Fourier transform of Gaussian function or Gaussian function, so the Gabor function has good time-frequency domain characteristic.The general type that complex value two-dimensional Gabor function space is expressed is

as follows

$h(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\exp (-\frac{1}{2}[{\left(\frac{x}{{\mathrm{\σ}}_x}\right)}^2+{\left(\frac{y}{{\mathrm{\σ}}_y}\right)}^2\left]\right)\exp \left(2\mathrm{\πjFx}\right)$

Wherein, F representes the centre frequency of Gabor function, σ _x, σ _yBe the scale factor of Gabor function, determined along the gaussian envelope of x and y axle.Therefore, (x y) is one and receives scale parameter σ h _x, σ _yThe complex value sine function of the Gaussian function modulation of decision.

As female ripple, it is carried out multi-frequency expansion and direction rotation with the Gabor function shown in the top formula, can produce the Gabor function of one group of self similarity:

$h^{\′}(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\exp (-\frac{1}{2}[{\left(\frac{x^{\′}}{{\mathrm{\σ}}_x}\right)}^2+{\left(\frac{y^{\′}}{{\mathrm{\σ}}_y}\right)}^2\left]\right)\exp \left(2\mathrm{\π}{F}^{\′}{x}^{\′}\right)$

When choosing the symmetrical Gabor wave filter of circle, i.e. σ _x=σ _yDuring=σ, get

p＝1，2，3，...S，q＝1，2，3，...L

x′＝xcosθ _q+ysinθ _q，y′＝-xsinθ _q+ycosθ _q

F′＝F _p， ${\mathrm{\θ}}_q=\frac{\mathrm{\π}(q-1)}{L}$

Wherein, integer index p, q have represented the value and the angle rotation of filter frequencies respectively, and S is the total degree that selecting frequency changes, and L is the total degree of direction rotation.

Centre frequency and scale factor rule of thumb satisfy relation:

$F=\frac{1}{2\mathrm{\σ}}$

So just formed with F and θ is the Gabor bank of filters of major parameter.

(2) make up the Adaptive Gabor bank of filters

Under off-line case, at first to the flawless reference image R of cloth (x y) carries out power spectrumanalysis, obtain the corresponding figures image texture the centre frequency group (F, F/2), F=0.16 wherein; Position angle group (θ, θ+pi/2), wherein θ=0.Set and form the Gabor bank of filters of S * L then with these two groups of parameters, the centre frequency number that the S representative is chosen, the position angle number that the L representative is chosen.At present embodiment, owing to chosen 2 centre frequencies and 2 position angles, thus S=2, L=2.

The Adaptive Gabor wave filter of embodiment of the invention structure can be expressed as: $h_{m,n}(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}\mathrm{Exp}(-\frac{1}{2}[{\left(\frac{x^{\′}}{\mathrm{\σ}}\right)}^2+{\left(\frac{y^{\′}}{\mathrm{\σ}}\right)}^2\left]\right)\mathrm{Exp}\left(2\mathrm{\π}{F}^{\′}{x}^{\′}\right),$ Wherein, the anglec of rotation x ' of coordinate axis x=xcos θ _n+ ysin θ _n, the anglec of rotation y ' of coordinate axis y=-xsin θ _n+ ycos θ _n, F '=F _m,

M=1,2, n=1,2, σ is the scale factor of Gabor function, has determined to revolve along a gaussian envelope angle of x and y axle;

3. off-line obtains the Defect Detection parameter

Under off-line case, (x y) carries out the filtering of Gabor bank of filters, promptly to normal cloth texture image R

H _m，n(x，y)＝R(x，y)*h _m，n(x，y)，1≤m≤S，1≤n≤L

Here, the * representative is a two-dimensional convolution, H _{M, n}(x y) is the feature image group of normal cloth texture image after the filtering.For each such feature image, calculate its average (Mea _{M, n}) and variance (Std _{M, n}), can be used as the threshold value of online detection split image flaw.

4. cloth flaw characteristic is cut apart

(1) the cloth flaw is online cuts apart

(x y) detects the feature image group H ' that obtains exporting with the Adaptive Gabor bank of filters of setting to detecting sample I _{M, n}(x, y):

H′ _m，n(x，y)＝I(x，y)*h _m，n(x，y) (4-1)

Use the average and the variance that obtain in the step and can be partitioned into visual flaw.Image after the Gabor bank of filters handled carries out thresholding to be handled:

Wherein τ is adjustable parameter, has determined the susceptibility of wave filter to Defect Detection, depends on testing environment to a great extent.Experiment shows that τ chooses 2-3.

(2) characteristic value normalization is handled

Because the flaw absolute feature value fluctuation range that obtains with average, variance preset threshold is very big; And be not easy unified tolerance; Also can merge better simultaneously for the result who makes different wave filters obtain; Need carry out normalization to the result and handle, adopt every bit to represent that with the relative mistake of they and view picture image averaging value relative mistake is defined as:

N _m，n(x，y)＝|F _m，n(x，y)-Mea _m，n|/Std _m，n (4-3)

Wherein, Mea _{M, n}And Std _{M, n}Be respectively feature image) average and variance.

After the absolute feature value is relative value through the normalization treatment conversion; Absolute feature value originally just is " mapped " between-1-1; So just can analyze and definite flaw and position thereof through this relative value easily, and be beneficial to follow-up eigenwert and relatively judge whether to exist flaw.

(3) Feature Fusion and binaryzation are cut apart

The image that thresholding is handled has embodied the shape facility that flaw demonstrates under different frequency and the different directions, and these images are superposeed to merge obtains final binary image, accomplishes Defect Detection.

Use image interfusion methods such as average fusion, weighted mean fusion can realize the fusion between image, use the fusion method of Bernoulli criterion among the present invention.It has strengthens big number, the advantage of reduction decimal.Image of the present invention merges and be divided into two steps: at first utilizing the Bernoulli criterion to merge the filtered of different orientations (can be referring to H.Sari-Sarrafand J.S.Goddard; " Vision systems for on-loom fabric inspection; " IEEE Transactions on Industry Applications; 35 (6): 1252-1259,1999), promptly obtain the fusion results under each centre frequency.Then, obtain fusion results after the fusion results of side frequency replacement number is average,, just can obtain the split image of flaw again through binary conversion treatment.

It is following to utilize the Bernoulli criterion to merge the step of filtered of different orientations:

1) carry out the fusion of L deflection with the Bernoulli rule:

$I_i(x,y)=\underset{j=1}{\overset{L}{\mathrm{\Σ}}}{N}_{i,j}(x,y)-\underset{l,k=1}{\overset{L}{\mathrm{\Σ}}}{N}_{i,l}(x,y)N_{i,k}(x,y)(l\≠k,1\≤i\≤S)$

2) calculate the fusion of S frequency image with arithmetic equal value

$I_{\mathrm{fusion}}(x,y)=\frac{1}{S}\underset{i=1}{\overset{S}{\mathrm{\Σ}}}{I}_i(x,y)$

Merge in the binary picture that obtains and often contain the spot that noise causes, influence detects effect, therefore need carry out denoising to binary picture.Concrete grammar is: count the size of minimum flaw, remove noise with this threshold value, select 60-80 in the experiment.Experiment shows that this method can eliminate noise speckle, accurately reflects the size of flaw.

(4) the flaw existence is judged

After the image binaryzation processing, the output background pixel value of system was 0 (deceiving), and target flaw pixel was 1 (in vain), and the coordinate position that calculate flaw this moment is easy to obtain.Concrete grammar is: to transversely add up line by line statistical pixel values and form curve of binary picture, promptly obtain the horizontal projection line; Along vertically adding up statistical pixel values and form curve line by line, promptly obtain the vertical projection line.Be in the position that the peak value section promptly is a flaw.

Can knit flaw image authentication whole algorithm process with disconnected, as shown in Figure 4, image (A) is the disconnected flaw image of knitting; Image (A1) and (A2) be normalized horizontal properties result after the Gabor filtering; Image (A3) is that (A4) is normalized vertical characteristic result after the Gabor filtering; Image (A5) is the fusion results on the centre frequency F1; Fusion results on image (A6) the centre frequency F2; Image (A7) is total fusion results; Image (A8) is the result after the binary conversion treatment.

Claims (2)

1. the cloth flaw detection method based on machine vision comprises the following steps:

(1) the normal cloth texture image of industrial camera collection P (x, y);

(2) (x y) utilizes power spectral-density analysis, obtains the centre frequency F and the azimuth angle theta of cloth image texture to P;

(3) two groups of parameters of definite figure image texture: the centre frequency group (F, F/2); Position angle group (θ, θ+pi/2) is utilized these two groups of parameter settings and is formed the Adaptive Gabor bank of filters of S * L, wherein, and the centre frequency number that the S representative is chosen, the position angle number that the L representative is chosen;

(4) under off-line case, (x y) carries out the filtering of Adaptive Gabor bank of filters, obtains characteristic image group H to normal cloth texture image P _{M, n}(x y), calculates the average (Mea of its every width of cloth figure _{M, n}) and variance (Std _{M, n}), as the threshold value that detects visual flaw.

(5) gather cloth image I to be detected (x, y);

(6) (x y) carries out filtering, obtains characteristic image group H to the cloth image I with the Adaptive Gabor bank of filters that makes up _{M, n}(x, y);

(7) use the average (Mea that obtains from (4) _{M, n}) and variance (Std _{M, n}) to H _{M, n}(x y) carries out the threshold value aftertreatment and obtains the absolute feature image sets

Wherein, τ is adjustable parameter, span 2-3;

(8) to image sets F _{M, n}(x y) carries out the normalization processing and obtains relative characteristic image sets N _{M, n}(x, y);

(9) to image sets N _{M, n}(x, the y) fusion of travel direction angle and frequency, and carry out binary conversion treatment and obtain detecting bianry image I _Bm(x, y);

(10) utilize the area information of minimum flaw to remove bianry image I _Bm(x, the noise in y) obtain final detection result I _Out(x, y);

(11) judge testing result I _Out(whether x y) contains flaw.

2. the cloth flaw detection method based on machine vision according to claim 1 is characterized in that, in the step (9), the fusion method of using the Bernoulli criterion is to image sets N _{M, n}(x, y) fusion of travel direction angle and frequency.

Images